What are some uses of template template parameters?

Question

I think you need to use template template syntax to pass a parameter whose type is a template dependent on another template like this:

template <template<class> class H, class S>
void f(const H<S> &value) {
}

Here, H is a template, but I wanted this function to deal with all specializations of H.

NOTE: I've been programming c++ for many years and have only needed this once. I find that it is a rarely needed feature (of course handy when you need it!).

I've been trying to think of good examples, and to be honest, most of the time this isn't necessary, but let's contrive an example. Let's pretend that std::vector doesn't have a typedef value_type.

So how would you write a function which can create variables of the right type for the vectors elements? This would work.

template <template<class, class> class V, class T, class A>
void f(V<T, A> &v) {
    // This can be "typename V<T, A>::value_type",
    // but we are pretending we don't have it

    T temp = v.back();
    v.pop_back();
    // Do some work on temp

    std::cout << temp << std::endl;
}

NOTE: std::vector has two template parameters, type, and allocator, so we had to accept both of them. Fortunately, because of type deduction, we won't need to write out the exact type explicitly.

which you can use like this:

f<std::vector, int>(v); // v is of type std::vector<int> using any allocator

or better yet, we can just use:

f(v); // everything is deduced, f can deal with a vector of any type!

UPDATE: Even this contrived example, while illustrative, is no longer an amazing example due to c++11 introducing auto. Now the same function can be written as:

template <class Cont>
void f(Cont &v) {

    auto temp = v.back();
    v.pop_back();
    // Do some work on temp

    std::cout << temp << std::endl;
}

which is how I'd prefer to write this type of code.

Actually, usecase for template template parameters is rather obvious. Once you learn that C++ stdlib has gaping hole of not defining stream output operators for standard container types, you would proceed to write something like:

template<typename T>
static inline std::ostream& operator<<(std::ostream& out, std::list<T> const& v)
{
    out << '[';
    if (!v.empty()) {
        for (typename std::list<T>::const_iterator i = v.begin(); ;) {
            out << *i;
            if (++i == v.end())
                break;
            out << ", ";
        }
    }
    out << ']';
    return out;
}

Then you'd figure out that code for vector is just the same, for forward_list is the same, actually, even for multitude of map types it's still just the same. Those template classes don't have anything in common except for meta-interface/protocol, and using template template parameter allows to capture the commonality in all of them. Before proceeding to write a template though, it's worth to check a reference to recall that sequence containers accept 2 template arguments - for value type and allocator. While allocator is defaulted, we still should account for its existence in our template operator<<:

template<template <typename, typename> class Container, class V, class A>
std::ostream& operator<<(std::ostream& out, Container<V, A> const& v)
...

Voila, that will work automagically for all present and future sequence containers adhering to the standard protocol. To add maps to the mix, it would take a peek at reference to note that they accept 4 template params, so we'd need another version of the operator<< above with 4-arg template template param. We'd also see that std:pair tries to be rendered with 2-arg operator<< for sequence types we defined previously, so we would provide a specialization just for std::pair.

Btw, with C+11 which allows variadic templates (and thus should allow variadic template template args), it would be possible to have single operator<< to rule them all. For example:

#include <iostream>
#include <vector>
#include <deque>
#include <list>

template<typename T, template<class,class...> class C, class... Args>
std::ostream& operator <<(std::ostream& os, const C<T,Args...>& objs)
{
    os << __PRETTY_FUNCTION__ << '\n';
    for (auto const& obj : objs)
        os << obj << ' ';
    return os;
}

int main()
{
    std::vector<float> vf { 1.1, 2.2, 3.3, 4.4 };
    std::cout << vf << '\n';

    std::list<char> lc { 'a', 'b', 'c', 'd' };
    std::cout << lc << '\n';

    std::deque<int> di { 1, 2, 3, 4 };
    std::cout << di << '\n';

    return 0;
}

Output

std::ostream &operator<<(std::ostream &, const C<T, Args...> &) [T = float, C = vector, Args = <std::__1::allocator<float>>]
1.1 2.2 3.3 4.4 
std::ostream &operator<<(std::ostream &, const C<T, Args...> &) [T = char, C = list, Args = <std::__1::allocator<char>>]
a b c d 
std::ostream &operator<<(std::ostream &, const C<T, Args...> &) [T = int, C = deque, Args = <std::__1::allocator<int>>]
1 2 3 4 

Here is a simple example taken from 'Modern C++ Design - Generic Programming and Design Patterns Applied' by Andrei Alexandrescu:

He uses a classes with template template parameters in order to implement the policy pattern:

// Library code
template <template <class> class CreationPolicy>
class WidgetManager : public CreationPolicy<Widget>
{
   ...
};

He explains: Typically, the host class already knows, or can easily deduce, the template argument of the policy class. In the example above, WidgetManager always manages objects of type Widget, so requiring the user to specify Widget again in the instantiation of CreationPolicy is redundant and potentially dangerous.In this case, library code can use template template parameters for specifying policies.

The effect is that the client code can use 'WidgetManager' in a more elegant way:

typedef WidgetManager<MyCreationPolicy> MyWidgetMgr;

Instead of the more cumbersome, and error prone way that a definition lacking template template arguments would have required:

typedef WidgetManager< MyCreationPolicy<Widget> > MyWidgetMgr;

Here's another practical example from my CUDA Convolutional neural network library. I have the following class template:

template <class T> class Tensor

which is actually implements n-dimensional matrices manipulation. There's also a child class template:

template <class T> class TensorGPU : public Tensor<T>

which implements the same functionality but in GPU. Both templates can work with all basic types, like float, double, int, etc And I also have a class template (simplified):

template <template <class> class TT, class T> class CLayerT: public Layer<TT<T> >
{
    TT<T> weights;
    TT<T> inputs;
    TT<int> connection_matrix;
}

The reason here to have template template syntax is because I can declare implementation of the class

class CLayerCuda: public CLayerT<TensorGPU, float>

which will have both weights and inputs of type float and on GPU, but connection_matrix will always be int, either on CPU (by specifying TT = Tensor) or on GPU (by specifying TT=TensorGPU).